Performance profiling + JVM tuning (ZGC, async-profiler)

Description

Before releasing to the public, the server JVM must be tuned for low-latency GC to avoid tick pauses that cause lag spikes during races. ZGC is the recommended GC for Minestom's workload.

Acceptance Criteria

• async-profiler run on a 20-player load test identifies the top 3 hot spots in the tick loop
• ZGC is configured and validated: no GC pause exceeds 5ms during a 20-player 15-minute session
• JVM startup flags documented in deploy/jvm-flags.sh and docs/ops/jvm-tuning.md
• Heap sizing recommendations documented (initial: 512MB, max: 2GB for 20 players)
• Off-heap direct memory usage checked (Minestom uses Netty direct buffers)
• Before/after benchmark numbers committed to docs/research/jvm-tuning-results.md

Technical Details

• Use java -XX:+UseZGC -XX:+ZGenerational for Java 25
• async-profiler: ./profiler.sh -d 60 -f out.html <pid>
• Key metrics: tick P50/P99/P999 duration, GC pause time, allocation rate

Dependencies

• "Load-test ECS game loop" (Alpha) provides the baseline metrics to compare against

Estimate

M

Start with the dependent “Load-test ECS game loop” baseline, then run the specified async-profiler command during a 20-player session and track tick P50/P99/P999, GC pauses, allocation rate, and direct memory. Review deploy/jvm-flags.sh, docs/ops/jvm-tuning.md, and docs/research/jvm-tuning-results.md. Done means validated ZGC results meet the 5ms pause target, recommendations and flags are documented, and before/after numbers are recorded.